Zhimeng Zhu

Incoming M.S. student in Electrical and Computer Engineering at Northeastern University.

Incoming M.S. candidate in Electrical and Computer Engineering at Northeastern University, beginning in Fall 2026, with prior automation training at Huazhong University of Science and Technology and research experience in computer vision, embedded AI, LLM-based systems, federated learning, medical imaging, and generative models.

happyrain0427@gmail.com 中文

Education

Northeastern University

2026-2028

Huazhong University of Science and Technology (HUST)

2022-2026

Publications

Enhancing Brain Tumor Detection: A Comparative Study of CNN Architectures Using MRI Data

2025

Zhu, Zhimeng. ITM Web of Conferences, 70, 03014.
DOI: 10.1051/itmconf/20257003014

Research Experiences

Bachelor's Thesis

05/2026

Cross-View Collaborative Target Recognition Based on Vision-Language Multimodal Models
Bachelor’s thesis, Huazhong University of Science and Technology. Advisor: Assoc. Prof. Tian Tian

  • Designed a lightweight cross-view collaborative recognition pipeline combining YOLO, CLIP, and BLIP for UAV multi-view observation scenarios.
  • Built semantic packet representations, cross-view semantic association, and multimodal fusion decision methods to improve recognition stability under viewpoint differences.
  • Implemented an experimental and interactive demo system, achieving 0.9775 F1 for cross-view association and 0.9527 F1 for fusion recognition, with about 577.9x communication compression compared with full-image transmission.

Multispectral Object Detection and Embedded Deployment

09/2024-present

National Key Laboratory of Multispectral Information Intelligent Processing Technology
Advisor: Assoc. Prof. Tian Tian, School of Artificial Intelligence and Automation, HUST

  • Designed a deep learning-based multispectral object detection model using TensorFlow and PyTorch to improve accuracy and robustness under complex lighting and environmental conditions.
  • Built a high-quality multispectral dataset through data collection, cleaning, and precise annotation for model training and validation.
  • Optimized model architecture and training strategies using data augmentation, residual connections, and batch normalization to reduce overfitting and vanishing gradient issues.
  • Deployed the optimized model on an NVIDIA Jetson Xavier NX platform with quantization and pruning for real-time embedded inference.

Intelligent Q&A System for Transportation Engineering Standards Based on DeepSeek

03/2024-05/2025

First Prize, 20th “Chuanshibao Cup” Transportation Science and Technology Competition
Advisor: Dr. Zehao Jiang, School of Civil and Hydraulic Engineering, HUST

  • Developed a domain-specific large language model for transportation engineering standards to address inefficient, inconsistent, and incomplete standard systems.
  • Processed national and industry standard documents through batch JSON structuring, data cleaning, and normalization.
  • Built on DeepSeek-R1 with 32B parameters and applied domain fine-tuning using a learning database and specialized lexicon, improving terminology recognition to F1 = 0.89.
  • Designed an output control mechanism for concise professional answers with automatic citation of standard references and section numbers.
  • Constructed a standards knowledge graph for conflict detection, redundancy identification, intelligent Q&A, and dynamic standards optimization.

Federated Learning for Personalized Financial Services

04/2024-04/2025

Project Leader, Key Laboratory of Image Processing and Intelligent Control
Advisor: Assoc. Prof. Lintao Ye, School of Artificial Intelligence and Automation, HUST

  • Applied federated learning to personalized financial services to improve recommendation accuracy and client risk assessment while preserving data privacy.
  • Reproduced FedAvg and FedProx and built a distributed PyTorch training framework with the Federated-Learning-PyTorch dataset.
  • Addressed Non-IID challenges such as cross-bank customer data, regional consumer behavior, and temporal transaction patterns through algorithmic tuning.
  • Integrated FedProx regularization to reduce client-side heterogeneity and improve model stability and generalization in imbalanced financial scenarios.

Enhancing Brain Tumor Detection Using CNN Architectures

07/2024-08/2024

Online Research Seminar: Machine Learning and Data Science - Development of Applications
Team Leader, Advisor: Prof. Mark Vogelsberger, Department of Physics, MIT

  • Proposed VGG19-BMT by adjusting convolutional layers and feature extraction modules for improved MRI-based brain tumor detection.
  • Preprocessed the Kaggle Brain MRI dataset with normalization and augmentation including rotation, brightness shift, and scaling.
  • Applied ReduceLROnPlateau, Dropout, and a custom classifier head to prevent overfitting on limited medical data.
  • Compared VGG19-BMT with VGG16, VGG19, ResNet18, and EfficientNet-B2 under consistent settings.
  • Achieved 99.51% accuracy and 0.9498 F1-score, outperforming all baselines.

Image Generation with Variational Autoencoders

06/2024-08/2024

Model Optimization and Improvement Exploration

  • Optimized the encoder, decoder, and custom loss function of a standard VAE and developed BetaVAE, Conditional VAE, and InfoVAE variants.
  • Preprocessed more than 200,000 CelebA face images with cleaning, normalization, rotation, flipping, and cropping.
  • Implemented all models in PyTorch and conducted GPU-based training with tuned learning rates, batch sizes, and optimizer settings.
  • Evaluated models using SSIM and user studies, identifying InfoVAE as the strongest variant with SSIM = 0.338.

Selected Projects

Cross-Domain Few-Shot Object Recognition

04/2025-06/2025

PyTorch, transfer learning, few-shot learning, domain adaptation

  • Implemented Prototypical Network, Improved ProtoNet, and Matching Network on the Office-Home dataset to explore cross-domain few-shot image recognition.
  • Applied target-domain fine-tuning to improve accuracy in cross-domain few-shot and 1-shot classification tasks.
  • Designed multi-source domain adaptation experiments to evaluate model performance under different domain shifts.
  • Showed that Improved ProtoNet achieved the best performance in most scenarios, especially in 1-shot tasks.

Autonomous Water Quality Monitoring System Based on Unmanned Surface Vehicle

01/2023-04/2023

Path planning, machine learning, statistical analysis, fuzzy mathematics

  • Designed an intelligent monitoring system integrating USV autonomous navigation with real-time data analysis for ecological protection and lake management.
  • Optimized USV path planning using A* and Dijkstra algorithms under complex lake topography and variable hydrodynamic conditions.
  • Implemented statistical analysis and machine learning methods to preprocess water quality data and build predictive models for environmental indicators.
  • Applied environmental science principles and fuzzy mathematics to improve model accuracy, robustness, and monitoring efficiency.

Additional Information

Computer Skills
Python, C, C++, MATLAB, Java, R

AI Frameworks
PyTorch, TensorFlow, DeepSeek-R1, federated learning frameworks

Research Methods
Computer vision, data preprocessing, model evaluation, domain adaptation, knowledge graphs

Interests
Piano, including 13 years of training and multiple awards with a provincial-level prize; basketball.